为什么gevent.socket会破坏multiprocessing.connection的认证？

Question

我有一个应用程序，它同时使用了 grequests 和 multiprocessing.managers，主要是为了实现进程间通信和异步的HTTP请求。

看起来，grequests 在使用 gevent.monkey 的 patch_all() 方法时，会破坏 multiprocessing.connection 模块，这个模块是 multiprocessing.manager.SyncManager 类及其衍生类所用的。

这个问题显然不是个别现象，而是影响到任何使用 multiprocessing.connection 的情况，比如 multiprocessing.pool 等等。

深入查看 gevent/monkey.py 的代码后，我发现将标准库中的 socket 模块替换成 gevent.socket 是导致问题的原因。这个替换可以在 gevent/monkey.py 的第115行找到，位于 patch_socket() 函数中：

def patch_socket(dns=True, aggressive=True):
    """Replace the standard socket object with gevent's cooperative sockets.
    ...
    _socket.socket = socket.socket # This line breaks multiprocessing.connection!
    ...

我的问题是，为什么这个替换会破坏 multiprocessing.connection，而使用 gevent.socket 相比于标准库的 socket 模块有什么好处？也就是说，如果我不替换 socket 模块，会有什么性能损失吗？

错误追踪信息

Traceback (most recent call last):
  File "clientWithGeventMonkeyPatch.py", line 49, in <module>
    client = GetClient(host, port, authkey)
  File "clientWithGeventMonkeyPatch.py", line 39, in GetClient
    client.connect()
  File "/usr/lib/python2.7/multiprocessing/managers.py", line 500, in connect
    conn = Client(self._address, authkey=self._authkey)
  File "/usr/lib/python2.7/multiprocessing/connection.py", line 175, in Client
    answer_challenge(c, authkey)
  File "/usr/lib/python2.7/multiprocessing/connection.py", line 414, in answer_challenge
    response = connection.recv_bytes(256)        # reject large message
IOError: [Errno 11] Resource temporarily unavailable

重现错误的代码

（在 Ubuntu 11.10 服务器上，使用 Python 2.7.3，并安装了 gevent、greenlet 和 grequests）

manager.py

## manager.py
import multiprocessing
import multiprocessing.managers
import datetime


class LocalManager(multiprocessing.managers.SyncManager):
    def __init__(self, *args, **kwargs):
        multiprocessing.managers.SyncManager.__init__(self, *args, **kwargs)
        self.__type__ = 'LocalManager'

def GetManager(host, port, authkey):
    def getdatetime():
        return '{}'.format(datetime.datetime.now())

    LocalManager.register('getdatetime', callable = getdatetime)
    manager = LocalManager(address = (host, port), authkey = authkey)
    manager.start()

    return manager

if __name__ == '__main__':
    # define our manager connection parameters
    port = 55555
    host = 'localhost'
    authkey = 'auth1234'

    # start a manager
    man = GetManager(host, port, authkey)

    # wait for user input to shut down
    raw_input('return to shutdown')
    man.shutdown()

client.py

## client.py -- this one works
import time
import multiprocessing.managers

class RemoteClient(multiprocessing.managers.SyncManager):
    def __init__(self, *args, **kwargs):
        multiprocessing.managers.SyncManager.__init__(self, *args, **kwargs)
        self.__type__ = 'RemoteClient'

def GetClient(host, port, authkey):
    RemoteClient.register('getdatetime')
    client = RemoteClient(address = (host, port), authkey = authkey)
    client.connect()
    return client

if __name__ == '__main__':
    # define our client connection parameters
    port = 55555
    host = 'localhost'
    authkey = 'auth1234'

    # start a manager
    client = GetClient(host, port, authkey)
    print 'connected', client
    print 'client.getdatetime()', client.getdatetime()
    # wait a couple of seconds, then do it again
    time.sleep(2)
    print 'client.getdatetime()', client.getdatetime()

    # exit...

clientWithGeventMonkeyPatch.py

## clientWithGeventMonkeyPatch.py -- breaks, depending on patch_all() parameters        
import time
import multiprocessing.managers


# this part is copied from grequests
# bear in mind that it doesn't actually do anything in this module.
try:
    import gevent
    from gevent import monkey as curious_george
    from gevent.pool import Pool
except ImportError:
    raise RuntimeError('Gevent is required for grequests.')

# this line causes breakage of the multiprocessing.manager connection auth method:
# Monkey-patch. 
# patch_all() parameters with default values:  socket=True, dns=True, time=True, select=True, thread=True, os=True, ssl=True, aggressive=True

curious_george.patch_all(thread=False, select=False) # breaks
#~ curious_george.patch_all(thread=False, select=False, socket = False) # works!
#~ curious_george.patch_all(thread=False, select=False, socket = True, aggressive = True, dns = True) # same as (thread=False, select=False); breaks
#~ curious_george.patch_all(thread=False, select=False, socket = True, aggressive = True, dns = False) # breaks
#~ curious_george.patch_all(thread=False, select=False, socket = True, aggressive = False, dns = True) # breaks
#~ curious_george.patch_all(thread=False, select=False, socket = True, aggressive = False, dns = False) # breaks







class RemoteClient(multiprocessing.managers.SyncManager):
    def __init__(self, *args, **kwargs):
        multiprocessing.managers.SyncManager.__init__(self, *args, **kwargs)
        self.__type__ = 'RemoteClient'

def GetClient(host, port, authkey):
    RemoteClient.register('getdatetime')
    client = RemoteClient(address = (host, port), authkey = authkey)
    client.connect()
    return client

if __name__ == '__main__':
    # define our client connection parameters
    port = 55555
    host = 'localhost'
    authkey = 'auth1234'

    # start a manager
    client = GetClient(host, port, authkey)
    print 'connected', client
    print 'client.getdatetime()', client.getdatetime()
    # wait a couple of seconds, then do it again
    time.sleep(2)
    print 'client.getdatetime()', client.getdatetime()

    # exit...

Answer 1

如果你不对socket模块进行修补，gevent就无法在网络操作时不被阻塞，这样你使用gevent的主要好处就没法享受了。

gevent和multiprocessing其实并不是特别兼容——gevent主要是希望你通过它来进行网络连接，而不是绕过Python的高级socket接口（而multiprocessing正是这么做的）。